Semantic Progress, Neural Alignment, and the Geometry of Shared Representations: A Candidate Framework for How Information Accumulation Constraints Are Shared Across Brains and Language Models

Abstract

This version corrects a wrong arXiv identifier. Version 2 cited the theoretical paper on chaotic regularization in recurrent neural networks, "Discrete signaling mediates chaotic regularization in recurrent neural networks", as arXiv:2606.04428 in six places, including item 3 of the abstract. That identifier belongs to an unrelated astrophysics paper; the correct identifier is arXiv:2606.04426, and all six citations are corrected. What the text says about the paper matches its abstract, so no claim changes. An internal reference in the selection criteria is reworded, typographic dashes are removed, the Greek letter eta now renders in the PDF, and the file carries a neutral name. The abstract below differs from version 2 only in that citation and in punctuation. The error was found by an automated check and confirmed by hand against arXiv; this version has not had a full claim-by-claim audit. A recurring structural pattern appears across recent work in computational neuroscience and natural language processing: the accumulation of new, non-redundant information over sequential inputs is governed by measurable geometric and information-theoretic constraints that appear to operate similarly, though not identically, in human neural systems and large language models (LLMs). This paper synthesizes six to seven findings from recent arXiv preprints (primary categories: q-bio.NC and cs.CL) to articulate a candidate framework we call semantic accumulation geometry: the hypothesis that the rate and structure of information gain over sequential inputs is constrained by representational geometry, and that this constraint produces measurable alignment between neural and model-derived accounts of meaning. Specifically, we draw on evidence that (1) LLMs selectively converge with human-shared neural semantic representations along specific semantic dimensions rather than globally arXiv:2606.11598; (2) semantic progress in multi-turn dialogue can be formalized as question-conditioned uncertainty reduction with tractable geometric structure arXiv:2606.12332; (3) chaotic recurrent networks produce smooth population codes through intrinsic regularization, providing a mechanistic substrate for stable representational geometry arXiv:2606.04426; (4) speech foundation models exhibit hierarchical brain alignment that is temporally structured rather than static arXiv:2606.02305; (5) hyperbolic geometry in hippocampal population activity may confer capacity and decoding advantages consistent with the demands of accumulating structured information arXiv:2606.10238; (6) the policy compression framework, when extended to include irreducible uncertainty costs, reframes cognitive cost as a joint function of complexity and conditional entropy arXiv:2606.13132; and (7) sparse autoencoder representations of neural activity obey optimality constraints that produce hierarchical feature structure arXiv:2606.02385. This is explicitly a heuristic reading, not a derivation: the sources share vocabulary and structural analogies but do not share a common formal apparatus. Falsification paths are named for each major claim. The framework is offered as a candidate structural pattern worth investigating, not a paradigm-level claim. Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted synthesis from an arXiv preprint corpus, originally drafted 2026-06-16, produced under the direction of Cristian Ruvalcaba, the accountable human author. Not peer-reviewed. AI disclosure. This work was produced with an agentic AI research apparatus operated by Saluca Labs. The apparatus drafted, searched and analysed under direction. Cristian Ruvalcaba is the human author and is accountable for the content. No AI system is listed as an author or contributor, because authorship entails accountability that a model cannot hold; this disclosure is the credit, and it is deliberately the whole of it.